Table 1.
Representative patent examples from the UAV technology dataset.
Table 2.
Dataset characteristics and technological category distributions.
Fig 1.
Contrastive pre-training architecture for multi-label patent representations.
The system processes patent abstracts through a shared encoder to generate embeddings that are optimized using multi-label contrastive objectives. The similarity computation considers both instance-level relationships and label co-occurrence patterns, enabling the model to learn representations that capture technological relationships and domain-specific semantic structures.
Fig 2.
Retrieval-augmented demonstration selection.
The system leverages contrastive embeddings to identify relevant patent demonstrations through multi-faceted similarity scoring. The retrieval process considers semantic similarity, technical domain alignment, and diversity constraints to select informative examples that guide multi-label classification decisions. Retrieved demonstrations are ordered to balance relevance and diversity, providing comprehensive coverage of the label space.
Fig 3.
The system combines retrieved demonstrations with embedding-guided attention to perform multi-label patent classification.
The prediction module employs decomposed inference for each category while considering inter-label dependencies, adaptive thresholding based on uncertainty, and prototype-based fallback for sparse categories. The integration of contrastive embeddings, demonstration patterns, and language model reasoning enables robust classification with minimal labeled examples.
Table 3.
Overall performance comparison under 5-shot setting.
Fig 4.
Overall performance comparison across evaluation metrics.
The proposed framework consistently outperforms baseline methods across Macro-F1, Micro-F1, LRAP, and Coverage Error metrics. Error bars represent standard deviations across 50 experimental episodes. Statistical significance indicators (***p < 0.001, **p < 0.01, *p < 0.05) show comparisons against our framework.
Fig 5.
Few-shot learning curves showing performance vs. number of shots.
Our framework demonstrates superior few-shot learning capabilities across all shot settings. Left: Macro-F1 performance prioritizes rare category detection. Right: Micro-F1 performance emphasizes overall classification accuracy. Error bars represent standard deviations across 50 episodes.
Table 4.
Computational efficiency comparison.
Table 5.
Ablation study: component contribution analysis.